/

CVE-2026-65090

Severity:

HIGH RISK

(7.8)

OS Command Injection (CWE-78) in NVIDIA NemoClaw NIM

CVE-2026-65090

Analyzed

Vulnerability type

Category:

OS Command Injection

CWE:

CWE-78

Impact:

Code Execution

Severity:

HIGH RISK

(7.8)

CVSS version:

3.1

Catch this CVE while code is being written

Detect vulnerable code directly in pull requests, so CVEs are fixed before merge.

Summary

NVIDIA NemoClaw for Linux contains an OS command injection vulnerability in its NIM management component. An attacker who can influence data processed by the component may cause unintended operating system commands to execute, potentially affecting confidentiality, integrity, and availability.

Affected context

The vulnerability occurs in the NIM management component of NVIDIA NemoClaw for Linux when externally influenced values are incorporated into operating system command execution. Risk depends on whether an attacker can reach the relevant management functionality and control data that is passed into a command-processing path.

Exploit preconditions

  • Attack surface: The NIM management component and its command execution functionality
  • Required access: Access sufficient to submit or influence data handled by the management component
  • Trigger point: Processing of attacker-influenced input during an operating system command operation
  • Why exploitation is possible: Input may reach a command interpreter or command-building routine without adequate separation, validation, or neutralization

Why this happens

The underlying bug is an OS command injection condition, mapped to CWE-78. This class of weakness occurs when software constructs or invokes operating system commands using externally influenced data without reliably separating data from command syntax or applying an appropriate allowlist. As a result, specially crafted input may alter the intended command and cause additional operations to be executed.

Potential impact

  • An attacker may execute code in the security context of the NemoClaw process
  • Sensitive information could be disclosed
  • Application or system data could be modified or deleted
  • The affected service may become unavailable
  • Compromised systems could be used to access adjacent resources, depending on permissions

General mitigation guidance

Mitigation generally focuses on removing shell interpretation from application-controlled operations and reducing the amount of untrusted input that reaches operating system interfaces.

Common defensive patterns include:

  • Prefer direct process-execution APIs that pass arguments separately rather than invoking a shell
  • Use strict allowlists for commands, options, and expected input formats
  • Apply contextual validation before values reach command construction
  • Avoid concatenating untrusted data into command strings
  • Run management components with least-privilege operating-system identities
  • Restrict filesystem, network, and process permissions available to the component
  • Treat administrative and management inputs as untrusted unless their provenance is explicitly enforced

How teams detect issues like this

Organizations can look for command injection risks through:

  • Static analysis rules identifying shell invocation, command-string concatenation, and untrusted data flowing into process APIs
  • Code review checks for dynamic command construction and insufficient input allowlisting
  • Runtime monitoring for unexpected child processes, shell interpreters, or anomalous command-line arguments
  • Application logging and policy enforcement around NIM management operations
  • Automated security review tools that trace data from external inputs to operating-system execution sinks

Sources

SourceLink
github.comgithub.com/NVIDIA/product-security/tree/main/2026/5872
nvd.nist.govnvd.nist.gov/vuln/detail/CVE-2026-65090
cve.orgcve.org/CVERecord?id=CVE-2026-65090

Frequently asked

questions

Exploitability depends on exposure of the NIM management component and whether an attacker can provide or influence input processed by its command execution path. It is not necessarily exploitable from an unauthenticated remote position in every deployment.

The attacker generally needs access to a management interface, workflow, or data source that can influence the vulnerable operation. The exact requirement depends on deployment configuration and access controls.

Avoid shell-based command construction, use direct execution APIs with separated arguments, and enforce strict allowlists and contextual validation for all externally influenced values.

Combine secure process-execution patterns with least privilege, restricted management-plane access, code review for command construction, static analysis, and monitoring for unexpected process activity.

Find out what's already
exploitable in your codebase.